Probabilistic Pose Estimation From Multiple Hypotheses

Abstract

Pose estimation assesses the 6D pose of one or many objects in a scene. Considerable attention has been dedicated to the advancement of pose estimation algorithms capable of identifying the orientation of multiple objects within a scene in cases where partial occlusion occurs. However, only a few works focus on developing a parallelizable hypotheses-based estimator that naturally handles object symmetries. These algorithms should also tackle some issues: meaningless perspectives, objects with multiple uncertain local poses but a single global correct pose, and multiple correct poses. This paper proposes a novel probabilistic algorithm for pose estimation that addresses these issues. This probabilistic algorithm combines the information from multiple cameras to achieve a unique prediction that assembles global object information. The algorithm is tested over synthetic objects that simulate these issues. It achieves a rotation error below 1.5°, and a translation error of 1.5 pixels in the datasets used. Those results suggest that the algorithm can handle the mentioned issues up to a certain accuracy. Additionally, the method is compared against a state-of-the-art methodology of the LineMOD dataset. This comparison shows that our algorithm can compete against state-of-the-art algorithms in terms of accuracy.

Keywords

• Convolutional neural networks
• deep learning
• graph neural networks
• pose estimation
• orientation estimation